Loading...
pricing
Every plan includes all drift detectors and blocking. Add team policies and shared learning when you need them.
solo
Full drift detection and blocking for individual developers.
team
popularYour agents get safer as your team ships.
enterprise
Deploy on your own network.
compare plans
| Feature | Solo | Team | Enterprise |
|---|---|---|---|
| Drift detection | Rule-based | Rule-based | Rule + LLM |
| Destructive action blocking | Alert only | Block + alert | Block + alert |
| Scope drift detection | ✓ | ✓ | ✓ |
| Retry loop detection | ✓ | ✓ | ✓ |
| Error spiral detection | ✓ | ✓ | ✓ |
| Cost runaway detection | ✓ | ✓ | ✓ |
| Desktop notifications | ✓ | ✓ | ✓ |
| Slack notifications | — | ✓ | ✓ |
| Webhook delivery | — | ✓ | ✓ |
| Session history | 7 days | 90 days | Custom |
| Supervised runs | 100 / mo | 10,000 / mo | Unlimited |
| Run replay | ✓ | ✓ | ✓ |
| Git diff capture | ✓ | ✓ | ✓ |
| Run comparison | — | ✓ | ✓ |
| Cost analytics | — | ✓ | ✓ |
| Shared drift patterns | — | ✓ | ✓ |
| Custom detection policies | — | — | ✓ |
| SSO / SAML | — | — | ✓ |
| RBAC | — | — | ✓ |
| Self-hosted deployment | — | — | ✓ |
| SLA | — | — | ✓ |
faq
Why can't I just review the diffs myself?
You can — on one run. But a typical developer runs 12+ agent sessions a day. Most diffs get merged without a careful read. Caret checks every tool call against your intent in real time, catching scope drift, destructive actions, and retry loops before they land in the PR.
How is this different from CI?
CI runs after the code is written and pushed. Caret runs while the agent is writing code. It blocks dangerous commands before they execute, detects drift in real time, and steers agents back on track — all before a PR even opens.
Does it slow down my agent?
No. Drift detection runs in under 10ms using pure rule-based checks. The hook handler completes well under the 200ms budget. Your agent won't notice.
What AI agents are supported?
Any agent that supports hooks or runs as a CLI process. Claude Code has native hook support. For Cursor, Aider, Codex, and others, Caret wraps them with `caret run` to capture and supervise.
Is my source code sent to Caret servers?
No. The supervisor runs entirely locally. Only the git diff (changes made during the run) is uploaded for replay. Your full source code never leaves your machine. On Enterprise, even diffs stay on your infrastructure.
What's the difference between watchdog and guardrail mode?
Watchdog mode (Solo) detects drift and alerts you via desktop notification. You decide what to do. Guardrail mode (Team) can automatically block destructive actions like `rm -rf`, force pushes, or DROP TABLE before they execute.
What is LLM-powered drift evaluation?
Enterprise includes an optional LLM evaluation step for ambiguous cases. When rule-based detectors flag a potential drift, the LLM evaluates whether the agent's action actually matches the original intent. This reduces false positives for complex tasks.
Can I self-host Caret?
Yes, on the Enterprise plan. You get the full platform — supervisor, dashboard, API — running in your own infrastructure. No data ever leaves your network.
How does billing work?
Team plans are billed monthly per seat. 14-day free trial, no credit card required. Enterprise plans are billed annually with custom terms.